Researchers have employed machine learning (ML) algorithms alongside firms' financial indicators to analyse ESG performance. The authors of [7] integrated annual financial indicators with natural language data in their study, utilising models such as XGBoost, CatBoost, a Keras sequential model, the K-Nearest Neighbours (KNN) classifier, and a simple linear regressor to predict firms' annual ESG ratings. The research conducted by [8] introduced a model for ESG score prediction by training a heterogeneous ensemble model that incorporates neural networks and the Random Forest algorithm. The authors of [9] not only examined the impact of financial indicators on ESG scores but also included governance factors to provide a more comprehensive assessment. Their study, which encompassed over 30 industries in Taiwan, including the oil, electricity, and gas sectors, tested four ML models: support vector machines (SVM), Random Forest, extreme learning machine (ELM), and extreme gradient boosting. Meanwhile, the authors of [10] initially implemented three standalone machine learning models: categorical gradient-boosted trees, feedforward neural networks, and gradient-boosted trees. They later proposed two ensemble models: a neural network ensemble and a heterogeneous ensemble combining all three individual models. Their findings revealed that the heterogeneous ensemble achieved the most accurate ESG predictions, as it leveraged the strengths of each particular model to uncover unique correlations between input variables and ESG scores. [3] investigated the external environmental factors and internal corporate attributes influencing ESG performance. They applied the Lasso method for data screening and employed ML models such as SVM, KNN, neural networks, and Random Forest to predict firms' ESG performance. Their study confirmed that both internal and external factors play a crucial role in enhancing corporate ESG performance. Furthermore, [11] and [12] incorporated social and governance datasets along with unstructured news and social media content, including Twitter data, to improve ESG score prediction. To analyse this unstructured text, they utilised natural language processing (NLP), which allowed the models to differentiate between pillar-specific input data and intermediate data representations. Their results highlighted the effectiveness of NLP in capturing the full scope of ESG-related data.

Existing literature has explored the use of financial indicators from firms in predicting ESG scores. In the study [7], various financial metrics were analysed, including the Return on Assets (ROA), Return on Equity (ROE), net income, total assets, price-to-free cash flow ratio, long-term debt, price-to-sales ratio, debt-to-equity ratio, quick ratio, earnings per share, current ratio, price-to-book ratio, price-to-earnings ratio, Earnings Before Interest, Taxes & EBITDA), return on investment, revenue, total liabilities, employee turnover rate, and the independence of the chair and Chief Executive Officer (CEO) roles. Additionally, [13] included the sales growth rate in their examination. The study [6] incorporated control variables, including the asset-liability ratio, Tobin's Q, operating cash flow, and firm size. Building on this, [14] added total return and dividend yield to the analysis. At the same time, [4] also considered a company's market value to investigate its relationship with ESG performance within the oil and gas industry.

According to [4], due to the considerable environmental impact of the oil and gas sector, ESG (Environmental, Social, and Governance) scores have gained increased relevance. This study examines the impact of ESG strategies on various financial variables, aiming to develop a comprehensive model of a firm's financial system. Specifically, it assesses the potential effects of each ESG pillar on a company's market value. The research utilises ESG metrics along with financial and economic data obtained from Refinitiv Eikon DataStream. To evaluate operational and financial performance, the study employs the ROA and ROE metrics, respectively. The researchers employed partial least squares structural equation modelling (PLS-SEM) to develop a new global ESG score and examine its impact on firms' financial performance. Their findings indicate that ESG practices have a positive effect on financial performance within the oil and gas industry.

The related study [6] examined how sustainability performance affects the Chinese crude oil market using a panel regression approach. One of the key findings is that crude oil futures prices have a significant influence on ESG performance, suggesting that price fluctuations can encourage companies to enhance their ESG ratings.

A reverse analysis of the ESG-finance relationship in the oil and gas industry, as conducted by [15], forecasted oil prices during the COVID-19 pandemic by examining stock markets, global environmental indices (ESG), and green energy resources (GER). Advanced ML algorithms, including Random Forest, LightGBM, CatBoost, XGBoost, and neural network models, were used in this study. The study [14] examined how the projected dividend growth and discount rates of Latin American (Latam) oil and gas companies are impacted by their corporate ESG performance. The researchers built their research upon the work conducted by Campbell and Shiller to compute the present value decomposition.

A noticeable gap exists in the existing literature, as limited studies have focused on predicting ESG scores in the oil and gas sector by integrating financial indicators, ESG metrics, and machine learning techniques. For instance, the research by [4] was limited to data from a single year (2020) and did not assess the predictive potential of their model or incorporate machine learning methods. Likewise, the studies [6] and [14] did not utilise machine learning in their analyses. While [15] did apply machine learning algorithms, ESG scores were merely used as inputs to analyse corporate financial performance, rather than being predicted. Moreover, past research has yet to establish a reliable prediction framework for ESG scores specifically within the oil and gas industry. As a result, investors and key decision-makers face difficulties in forecasting firms' ESG performance.

The numerical results obtained from models utilising regression machine learning algorithms were evaluated based on their prediction accuracy. In the research reported in [7], [8], and [10], researchers used the mean average error (MAE) and the coefficient of determination (R²) parameters. In the study [9], two additional performance variables were introduced: root mean square error (RMSE) and mean absolute percentage error (MAPE), for comparison of the models' results. The study [11] used the RMSE and MAE to examine the models obtained for predicting ESG scores.

This report examines the performance of ESG scores through the lens of stakeholder theory. ESG scores and investments offer insights into various aspects of the company, including its internal governance competencies, interactions with local governments, and other social responsibilities. Furthermore, companies are inclined to embrace social responsibilities, including fulfilling debt obligations [16].

Previous research has shown that within areas featuring well-established ESG scoring systems, investors incorporate high ESG performance as a key factor in making investment decisions [17]. Additionally, stakeholders, as suggested by [18], believe that companies which prioritise strong ESG practices and a positive reputation are more likely to compete effectively in the market. Consequently, successful management, recognised by stakeholders, can lead to increased investment, productivity, and consumption.

Furthermore, according to [19], firms can improve their financial performance through the indirect benefits of maintaining solid stakeholder relationships. Companies that engage in socially and environmentally responsible practices, coupled with effective governance processes, align with the interests of stakeholders, thereby enhancing their financial performance. Furthermore, the resource-based hypothesis suggests that businesses with heightened ESG disclosure (ESGD) can gain competitive advantages, and that the financial performance appears to be positively correlated to ESGD. This suggestion indicates that companies can strengthen their corporate reputation and develop internal resources by engaging in ESGD initiatives.

Additionally, [7] demonstrated that financial ratios, such as ROA and ROE, can be accurately predicted using machine learning models based on ESG ratings, thereby reinforcing the strong link between ESG performance and financial outcomes. Similarly, [3] applied machine learning methods to fuse macroeconomic and microeconomic data, revealing that internal financial indicators play a significant role in shaping corporate ESG performance. In parallel, [12] established that strong financial results are positively associated with the adoption of sustainable business models, demonstrating that companies with solid financial standing tend to achieve better ESG outcomes while simultaneously reducing investment risk and securing competitive returns.

LSEG Data & Analytics (LSEG) is a globally recognised provider of financial data, operating in 175 countries. Beyond its role as a key source of financial information, numerous research studies have analysed its ESG disclosures, which encompass data from over 12,000 public and private companies, covering more than 80% of global market capitalisation. The methodology used to determine sustainability ratings is entirely data-driven, carefully adjusted to reflect industry-specific criteria, and refined to mitigate biases related to market capitalisation and transparency [20]. For companies in the oil and gas industry, the parameters used in ESG score calculations were compared. The environmental (E) and social (S) pillar scores are assessed in relation to firms and their industry counterparts, whereas the governance (G) pillar score is evaluated based on the country of incorporation. The total ESG score is derived from the combined scores of these three pillars. These scores are not absolute but are instead calculated relative to the broader industry and market context in which the company operates [21]. A total of 186 selected metrics contribute to the ESG scores, which are further categorised into 10 groups, as illustrated in Figure 1. Consequently, the overall scores provided by Refinitiv are utilised for this analysis, in alignment with the methodologies applied by [10], [8], and [4].

Figure 1.

ESG score computation

The input variables, consisting of historical financial indicators relevant to firms in the oil and gas industry, are also obtained from LSEG. The retrieval of both financial indicators and ESG scores was carried out using two separate methodologies, Datastream and Refinitiv, both of which are well-established data sourcing methods provided by LSEG.

The data were extracted from the detailed examination of the literature. Notable differences can be observed between the financial indicators derived using the Datastream variables method and those selected via Refinitiv. These variations are primarily due to the differing data sources and the unique calculation methodologies applied to each indicator by the respective platforms. Appendix Table A1 outlines these discrepancies and highlights the indicators that were excluded from the analysis when Refinitiv was used as the data source.

The dataset comprised companies from the oil and gas industry that had reported their ESG scores between 2010 and 2022. This selection process resulted in a total of 497 companies, whose data were acquired using both the Datastream and Refinitiv methodologies. Table 1 provides a breakdown of the number of companies from each continent included in this analysis.

Over the past decade and a half, the incorporation of ESG scores and related indicators into financial reporting has steadily increased within the oil and gas industry. To gain deeper insights into ESG score trends and patterns, a large sample size spanning 12 years, ranging from 2010 to 2022, was chosen for analysis. Machine learning algorithms are particularly effective in predicting outcomes from extensive datasets, as they can identify intricate and hidden relationships while improving prediction accuracy. Therefore, the comprehensive dataset covering 12 years is well-suited for ML analysis. The exclusion of ESG scores for 2023 was due to incomplete disclosures for the selected companies. Since ESG scores serve as the primary predictor, missing data were not processed to prevent introducing noise into the dataset, as such inconsistencies could compromise the accuracy and reliability of the prediction model [22].

Dummy variables were integrated into the dataset alongside ESG and financial indicators to enable a more detailed analysis. The explanatory variables considered in the study include country, firm size, and ESG scores. The purpose of dummy variables is to examine how firm size and geographic location influence ESG scores.

Before applying machine learning algorithms, the extracted data from both sources required preprocessing. The following sections describe the preprocessing steps.

This research employed panel data analysis, combining both cross-sectional and time-series dimensions. This method allows the modelling of unobserved heterogeneity and captures consistent characteristics over time, such as economic behaviour or management practices [29]. Following [30] and [31], a one-year lag between ESG scores and financial indicators was used to support this approach.

Data analysis diagram.

The data analysis process followed in this study is schematically shown in Figure 2. The process began with data collection, followed by multiple stages of data preprocessing. After preprocessing, the data were processed using two separate platforms, MATLAB and Azure AI/Machine Learning Studio, where ML algorithms were applied. The dataset was automatically split into training and testing subsets based on the default specifications of each platform. No parameter tuning was performed.

Figure 2.

Data analysis diagram

Subsets.

The datasets were organised into seven distinct groups. Figure 3 and Figure 4 illustrate the final sets used for both regression and classification ML analysis.

Dummy variables were integrated into the dataset alongside ESG scores and financial indicators. The explanatory variables selected for analysis included country, firm size, and ESG scores. The inclusion of dummy variables aimed to enable a comprehensive evaluation of how firm size and country influence ESG scores. Additionally, for datasets that produced the most promising results, further analyses were conducted. These included testing models without dummy variables, removing missing financial indicators, and assessing their impact, as detailed in Table 4. In total, six data sets were tested in this analysis, as shown in Figure 3 and Figure 4.

Figure 3.

The four Refinitiv data sets

Figure 4.

The two Datastream data sets

Table 4.

Financial indicators with high missing values from Refinitiv

Indicator	Missing values (out of 2834)
Total Assets Actual	901
Total Liabilities	601
Return On Assets Actual	692
Tobin's Q	2514
Price to FCF per Share	1239
Price to Sales Ratio Mean	2493

The seven data sets were fed into regression learner analysis, using continuous ESG scores. The obtained R² values of Refinitiv 1 are provided in Table 5. The highest value of R² was found to be 0.892, obtained from the Bagged Tree (ensemble) model. R² is used to evaluate the effectiveness and fit of regression models; the higher the R² value (the closer to 1), the greater the explanation of the variance in the dependent variable (ESG scores). This is explained by the independent variables (financial indicators and ESG scores), according to [34]. Generally, a model with an R² value greater than 0.9 is considered reliable; hence, a value of 0.892 is slightly lower than being reliable.

To test the effect of companies with missing ESG scores, Refinitiv 2 was evaluated, and the R² values obtained were compared to those of Refinitiv 1, as shown in Figure 5. R² value was reduced from 0.892 to 0.812 for the ensemble Bagged Tree.

Figure 5.

Comparison between R² values of Refinitiv 1 and 2 (encountered missing ESG scores)

In addition to evaluating the models based on the R² value, other performance metrics, such as MAE, MSE, and RMSE, were also analysed. Table 6 and Figure 6 present the RMSE, MSE and MAE values for both Refinitiv 1 and 2. According to [35], MAE and MSE are used to evaluate models by considering the distance between the predicted values and the actual training data points. MSE is more sensitive to outliers than MAE. The values obtained from the five models are promising, as the error values are small (the maximum error obtained is 0.0617 for MAE and 0.00717 for MSE). The ensemble Bagged Tree model outperforms the other two models, with an MAE value of 0.0535 and an MSE value of 0.00571. Furthermore, RMSE is commonly used to standardise the units of MSE. The distinct types of regularisation inherent to these metrics influence their relative effectiveness depending on the structure of the data [35]. The ensemble Bagged Tree outperforms other models, as it has the lowest RMSE value.

The values of MAE, MSE, and RMSE are zero if the regression model fits the data perfectly, and they are positive if the fit is not perfect. Therefore, models with lower values (closer to 0) exhibit better prediction performance [35]; accordingly, the error results obtained from the regression analysis of this dataset are excellent.

Figure 6.

Comparison between RMSE, MSE, and MAE values for Refinitiv 1 and 2

Figure 7.

Comparison between R² values for Refinitiv 1 and 3 (no missing financial indicators)

Figure 7 presents a comparison of R² values for Refinitiv 1 and 3, testing the effect of removing the missing financial indicators from the dataset. The R² value was 0.892 before exclusion and 0.858 after exclusion.

Figure 8 presents a comparison between Refinitiv 1 and 4. It is observed that once the dummy variables are removed from the dataset, the performance of the regression analysis improves slightly in terms of R², from the highest accuracy of 0.892 obtained from the Bagged Tree in Refinitiv 1 to 0.920 obtained by the linear regression in Refinitiv 4 (an increase of 2.779%). This result introduced and made the robust linear regression models superior to the ensemble Bagged Tree, which obtained the highest accuracy. One possible explanation for the increase in performance after the dummy variables are removed is that the linear regression exhibits limitations in capturing the nonlinearity and complexity inherent in the data system with dummy variables [36]. Therefore, as a result of having complex sets with dummy variables, linearity is not assumed, so tree-based algorithms are more applicable than linear regression [33]

Figure 8.

Comparison between R² values for Refinitiv 1 and 4 (no dummy variables)

Table 7 presents a comparison between Refinitiv 3 and 4 performance metrics in terms of the values of MAE, MSE, and RMSE. A similar regression analysis was performed on the data obtained from Datastream, and the results obtained are presented in Figure 9. The highest R² value of 0.878 was obtained from the Bagged Tree, compared to 0.892 from Refinitiv regression analysis.

Figure 9.

R² values of Datastream 1 (no missing ESG scores set)

To test the effect of utilising the full data set with missing ESG scores, Datastream 2 was used, and the values of R² obtained are compared to those of Datastream 1, as provided in Figure 10. The value of R² for the ensemble Bagged Tree increased slightly from 0.878 to 0.888 when the entire data set with missing values was used. The results indicate that sample size is crucial in machine learning. Hence, both classification and regression analysis would significantly benefit from a larger set of companies and historical data on ESG and financial indicators extended for a decade or more, along with robust features [7].

Table 8, Figure 10, and Figure 11 represent the performance metrics of both Datastream 1 and 2 datasets for the regression analysis.

Figure 10.

Comparison between R² values of Datastream 1 and 2 (encountered missing ESG scores)

Figure 11.

Comparison between RMSE, MSE, and MAE values of Datastream 1 and 2

Regression was applied to the Refinitiv 1 dataset, and the results are provided in Table 9. The highest R² value was obtained from the Voting Ensemble algorithm, at 0.922, followed by slightly reduced values from the other algorithms: Extreme Random Trees (0.920), Elastic Net (0.919), Random Forest (0.917), LightGBM (0.914), and XGBoost Regressor (0.906). Since a greater R² value suggests a better fit for the data set, its value is considered important. A value closer to 1 is aimed for, as when the R² core is 1, the regression predictions and the data are perfectly matched [37].

As explained by the results obtained from MATLAB, the closer the values of MAE and RMSE to zero, the better the model's performance. According to the obtained results, the highest MAE and RMSE values among the six models were 0.0516 and 0.0701, respectively, obtained by the XGBoost Regressor. Those values, although the highest, are promising and close to the error values of the other five models.

When relative variances in a regression job are more important than absolute values, the MAPE, which highlights percentage errors, is a suitable metric [35]. The values of MAPE obtained from this analysis showed the lowest value when the Elastic Net was used, with a percentage of 18.725, followed by Random Forest (19.455), Voting Ensemble (20.059), LightGBM (20.576), and XGBoost Regressor (22.971).

Testing the effect of removing the financial indicators, Table 10 represents a comparison between the R² values of Refinitiv 1, 2, and 3. The results show a slight reduction in the R² value for all the models, except for the XGBoost Regressor, whose R² value increased from 0.907 to 0.913 after the removal of the missing financial indicators. The influence of reducing the sample size in regression analysis on the performance of R² is consistent with the findings from MATLAB.

Figure 12 illustrates the comparison of R² between Refinitiv 1, 3, and 4 for the regression analysis data set. First, the results of Refinitiv 4 show that the values of R² obtained for the models Voting Ensemble, Extreme Random Trees, Elastic Net, Random Forest, and LightGBM were slightly lower compared to the values obtained from Refinitiv 1. Using dummy variables in the dataset had a very slight negative impact on accuracy when the Azure platform was employed. Table 11 and Table 12 represent the obtained performance metrics for Refinitiv 1, 3, and 4, indicated by their values of MAE, RMSE, and MAPE.

Figure 12.

Comparison between R² values of Refinitiv 1, 3, and 4

Figure 13, Figure 14, and Figure 15 illustrate the feature importance of country level, firm size, and ESG scores for the Refinitiv 1 dataset, respectively. The minimum-redundancy maximum-relevancy (mRMR) feature selection method was applied in this analysis, as it has been shown to outperform traditional top-ranking methods in various recent studies, as noted by [38].

Figure 13.

Importance score country level

Figure 14.

Importance scores firm size

Figure 15.

Importance scores ESG score

The results obtained from Bagged, Boosted, and Decision Tree models for the Refinitiv 1 dataset were closely aligned. Each algorithm yielded high accuracy but emphasised different factors. In the case of the ensemble Boosted Tree model, one possible reason for its performance is that boosting operates sequentially, correcting errors from previous trees to minimise bias in the final prediction. The ensemble's final output is derived from the collective predictions of all trees in the model [39]. Decision Trees are widely recognised for their effectiveness in regression analysis [40], as demonstrated in the results across the datasets of Refinitiv 1, 2, 3, and 4. The strong performance of these models may be attributed to the efficiency of the tree traversal algorithm [40] and their high generalisation capability [25] Regarding ensemble Bagged Trees, their effectiveness lies in the fact that bagging generates multiple subsets of the original dataset and trains them independently [41] This approach utilises all available dataset features and integrates them simultaneously to produce a collective prediction [42]. Linear regression performed best for Refinitiv 4 when the dummy variables were removed from the data set. For the results obtained in MATLAB, linear regression failed to explain the results, except for this data set. The highest R² value and the lowest MAE, MSE, and RMSE values for Refinitiv 4 were obtained from the robust linear regression (a form of linear regression algorithms that utilises weighted least squares, including the weight as anadditional scaling factor in the fitting process), which improves the fit. The weights determine how much each response value affects the final parameter estimates, with low-quality data points, such as outliers, having less influence on the fit [38].

The Voting Ensemble delivered the highest performance among the regression models, though its advantage over other models was only marginal. The strong performance of the Voting Ensemble in machine learning tasks is attributed to its ability to combine multiple learning models rather than relying on a single predictor. It generates multiple decision trees and determines the preferred output class through a majority vote [43]. XGBoost was employed for regression analysis due to its advantages, including high accuracy, sensitivity, and precision. The model achieved a regression score of 0.906, indicating a reliable and well-performing outcome. The results obtained from the Random Forest algorithm were also reliable, likely due to its robustness against noise and resistance to overfitting [25]. Both Random Forest and LightGBM are well-suited for regression problems, effectively handling continuous values [44]. However, the Random Forest was unable to interpret the results for the Refinitiv 3 dataset accurately. Regarding the Extreme Random Trees algorithm, its results were very similar to those of other models, demonstrating its computational efficiency. According to [45], using all original training samples instead of bootstrap replicas in the randomised tree strategy helps reduce bias, thereby contributing to the model's effectiveness.

The Elastic Net algorithm, which is built upon traditional regression or logistic regression [46], was utilised for the regression analysis. The obtained R² values were also close to those obtained from the regression results of Voting Ensemble, Extreme Random Trees, Random Forest, LightGBM, XGBoost Regressor, and Decision Tree. However, the Elastic Net model achieved the lowest MAPE amongst the other algorithms. The performance of this model is attributed to its ability to assess the random effects of panel data regressions that incorporate multiple interactions [21]. Lastly, in Azure machine learning, the Decision Tree could interpret the results of the regression analysis of datasets Refinitiv 3 and 4 only, meaning that the Decision Tree was only applicable with the reduced datasets. This situation could be attributed to the fact that Decision Tree utilises a divide-and-conquer strategy, excelling when there are only a few highly relevant attributes. However, their performance declines when there are numerous complex interactions among attributes. Errors can propagate through the tree, especially as the number of classes increases. Additionally, as the tree grows larger, the records within leaf nodes often become too small to support statistically significant decisions [25]. Furthermore, another possible reason is that the Decision Tree could be unstable, and the difficulty in managing the tree size [40]

The results obtained from the Datastream show a significant reduction in performance compared to the Refinitiv-sourced data. A comparison between the Datastream and Refinitiv regressions reveals that the highest R² value of 0.878 was obtained from the Bagged Tree, while the Refinitiv regression analysis yielded an R² of 0.892. Furthermore, the value of R² for the ensemble Bagged Tree increased slightly from 0.878 to 0.888 between Datastream 1 and 2. This reduction in performance for Datastream-sourced data could be attributed to overfitting; the sample dataset obtained from Datastream is too small compared to Refinitiv, and as a result, inadequate data are provided for training and testing the machine learning algorithms. According to [32], the more training data provided, the stronger the generalisation ability. The volume of training data available significantly influences the accuracy of learning outcomes, and improved performance correlates with larger datasets. Consequently, addressing data shortage becomes imperative to mitigate potential overfitting during training, thereby preserving the model's generalisation capability [32]

At the country level, this study examined the impact of country-level sustainability performance. Feature importance analysis revealed that country-level factors had the strongest influence. Notably, Malaysia emerged as the most significant indicator, with a feature importance value of 0.934. Malaysia has been a focal point in ESG research since implementing its first corporate social responsibility (CSR) framework in 2006. The integration of CSR into the Tenth Malaysia Plan, along with mandatory CSR disclosure requirements under the Companies Act 2016, has contributed to high CSR scores among Malaysian firms. As a result, Malaysia has positioned itself as a global leader in CSR reporting, with approximately 97% of its top 100 companies disclosing their corporate sustainability performance [47]. These findings align with [48], who argued that countries with strong ESG improvement initiatives often experience positive economic outcomes from their sustainability efforts. Furthermore, this effect is more pronounced in nations with supportive regulatory, legal, financial, and cultural frameworks for ESG implementation.

At the firm level, feature importance was found to be lower than at the country level. Mega-capitalisation companies had the highest firm-level feature importance, with a value of 0.131. Company size plays a crucial role in determining access to ESG initiatives and financial stability [31]. Several factors underscore the importance of firm size in the relationship between ESG and financial performance. Larger firms generally possess greater financial resources than smaller ones, allowing them to invest more extensively in sustainability initiatives [10]. Additionally, large corporations typically have structured strategies and business monitoring mechanisms that facilitate effective sustainability management. Moreover, firm visibility plays a key role, as prominent companies are more likely to adopt superior sustainability practices to maintain their reputation among stakeholders [49]

Furthermore, the substantial difference in feature importance scores between the highest-ranked and second-ranked features suggests a high level of confidence in the feature selection process. This significant drop indicates that MATLAB is highly confident in identifying the most critical predictor [38].

Most studies highlighted in this work suggest a positive relationship between ESG performance and financial performance [10]. The results obtained from this study are aligned with the results obtained from [8] (where MAE = 11.2 and R² = 0.54) and [9] (see Table 13), which relied on quantitative approaches utilising firms' financial indicators to predict ESG scores, and are superior to the results of [10] (where MAE = 1.22 and R² = 0.05) and [11] (see Table 14). Furthermore, this research addresses the gap in the literature where the prediction of ESG scores, specifically in the oil and gas industry, utilising machine learning and firm financial indicators, is lacking.

According to stakeholder theory, adopting ESG practices and promoting transparency can contribute to enhancing a company's overall value [50]. Engaging with stakeholders and addressing their concerns through socially responsible actions can help mitigate externalities and build stronger, positive relationships. As a result, an increasing number of companies are now publicly disclosing their ethical, social, and environmental practices and actively working to reduce their negative environmental impacts [51]. The authors of [52] argue that ESG initiatives can create shareholder value by boosting firm cash flows; for instance, a strong corporate reputation can drive higher product sales, and skilled employees can improve productivity. In comparison, the authors of [10] emphasise that sustainability-oriented firms offer greater utility to socially conscious investors. These insights suggest that sustainability initiatives align more closely with the interests of stakeholders. Furthermore, firms exhibiting stronger financial performance are believed to have a greater capacity to undertake social responsibilities, as higher profitability enables more significant investment in socially responsible activities [51].

In particular, the oil and gas sector faces mounting ESG pressures from investors, regulatory bodies, and the broader public. Consequently, ESG scores have become an increasingly critical factor in investment decision-making, with their significance continuing to rise each year. In response to these trends, this research presents a comprehensive methodology designed to assist stakeholders and policymakers in predicting ESG scores, thereby supporting more sustainable investment decisions [9].